Quadro only allowed multi-GPU in 0.8.1?

I just read the following from the 0.8.1 Programming Guide:

The 0.8 guide says nothing of the sort. Is this really going to be the case starting in 0.8.1?

We also noticed that new paragraph. It would be nice if the GeForce 8800GTX was “grandfathered in”, or at least unofficially allowed to work until the final 1.0 release of CUDA so that developers already using multiple GeForce 8800GTX cards could continue their work through a couple more beta versions. There may be very good reasons not to support them officially that we are unaware of, but allowing them to “eek along” with the present level of capability until the official “1.0” release would be very helpful for those of us doing initial development and testing on them.


John Stone

Wow, I didn’t see that change either. We are planning to build a dual 8800 GTX system when the 64-bit drivers come out, but we don’t have the budget for two Quadro cards (8x the price!). Is there a technical reason to limit the non-Quadro cards to one CUDA device per system?

Even if the GeForce cards are capable of multi-GPU only up to the “1.0” release, I’d be extremely disappointed. The Quadro 5600 is 5-6x the MSRP of the 8800GTX - which basically ruins the cost side of my business case for using CUDA to replace our cluster.

It’s reasonable to guess this is a sales/marketing decision and not a technical one considering John is already using the 8800GTX in a multi-GPU configuration.

I don’t know that my voice would make much difference, but I’ll express my opposition and disappointment anyway. To all the nVidia employees that frequent this board: you’ve been gracious enough to provide the SDK, samples, advice for performance improvement, help with programming issues, and so on to encourage CUDA adoption. I’d say CUDA has built up significant momentum. Please pass the word up the food chain that limiting multi-GPU CUDA to Quadros is a bad idea; it would only slow the momentum you’ve worked to create, IMHO.


I was going to wait for a clarification/explanation from an NVIDIA employee before drawing any unhappy conclusions, but, yes, that is one possibility.

Just to keep this near the top: When I read it in the new manual I though it was just a mistake - they have not actually changed anything yet as the 0.81 release was just the SDK - samples, whitepapers and doco - base tools have not changed. Next release in May with 8600 support is the first opportunity to introduce this.

If it’s some artificial limit built into the driver/CUDA library, and not the hardware, then I’m sure it’s possible to get around it. It’s just “officially not supported”.

I am almost sure that there is no technical reason for it, given that the GF will surely support SLI for the gamer community. So if two cards can work together, why should they not work separately?

CUDA is (currently) using the graphics driver. As long as there is no special low-level device driver, I don’t see any reason on the software side either that should prevent using multiple cards in CUDA, as we will always be able to use multiple cards for multiple graphics contexts. So my conclusion is indeed this is a marketing trick.

NVIDIA engineers, don’t let them turn your whitepapers into a black book.


This is extremely disappointing … It really blows many projections we had made based on the 8800 GTX prices out. Comeon NVIDIA don’t Celeronize the card.

Let the NVIDIA name be associated with performance and not crippleware

I was also quite disheartened when I read that paragraph.

Can we get an official ruling from the nvidia folks on this?

One other question I had in regards to the new Multi-GPU text in the CUDA programming guide is whether or not one could use a GeForce 8800GTX and a Quadro 4600 in multi-GPU mode, or if one can only use more than one card if all of them are Quadros.

John Stone

The 0.8.1 manual included text stating that multi-GPU operation is not available on GeForce products.
After reviewing this decision, multi-GPU operation will continue to be available on GeForce 8-Series products.
Configurations with multiple GPUs running CUDA must be of the same board type.

Thank you very much. Can you please let us know what you mean with same board type.

Configurations that have multiple instances of the same board: for example 2 GeForce 8800 GTX.

Thank you, Massimiliano and anyone else, for reconsidering. This is good news.

Jim Hardwick

Yes, this is indeed great news. Three cheers from all of us here!

John Stone

Thanks, this is good news for us as well!


Great to hear! Now we can build our dual GPU compute nodes as planned in May.

Cheers guys! Thanks to NVIDIA for reconsidering :hug: